Purchase Intentions and Purchase Behavior

Author(s): Donald G. Morrison

Source: Journal of Marketing, Vol. 43, No. 2 (Spring, 1979), pp. 65-74
Published by: American Marketing Association
Stable URL: http://www.jstor.org/stable/1250742 .
Accessed: 23/05/2014 10:35

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

American Marketing Association is collaborating with JSTOR to digitize, preserve and extend access to
Journal of Marketing.

http://www.jstor.org

This content downloaded from 62.197.243.108 on Fri, 23 May 2014 10:35:51 AM

All use subject to JSTOR Terms and Conditions
 G.MORRISON
DONALD
This paper provides a frameworkfor collecting, analyzing, and inter-
preting purchase intentions data.
PURCHASE
INTENTIONS
AND PURCHASE
BEHAVIOR
ANY surveyscontainpurchaseintentionsques-
M tions on such items as new food products, fre-
quently purchased package goods, appliances, au-
tomobiles, and capitalequipment.The time framemay
range from one week to 24 monthsor more. Although
these purchaseintentions data are used, there has not
been much work done on evaluating the accuracy of
these data-neither at the individual nor aggregate
levels. In particular,there has been little follow-up to
see if individuals surveyed actually purchased the
productof interestover the specified time periods. One
purpose of this paper is to stimulate such follow-up
procedures.
Purchase intentions continue to be an important
concept in marketing.The publishedliteraturecontains
a very small fractionof the actual studies which have
used purchaseintentions;nevertheless, the quantityof
literatureis quite large. Axelrod (1968) analyzes pur-
chase intentions along with other attitude measures
used for predicting actual purchase behavior. Smith
DonaldMorrisonis Professorof Businessat ColumbiaUniver-
sity, New York.The workreportedhere was supportedby the
Centerfor FoodPolicyStudies at the GraduateSchoolof Busi-
ness, ColumbiaUniversity.
Journal of Marketing
Vol. 43 (Spring 1979), 65-74.
This content downloaded from 62.197.243.108 on Fri, 23 May 2014 10:35:51 AM
All use subject to JSTOR Terms and Conditions
(1965) presents a highly readableand interestingdis-
cussion on the use of purchaseintentionsin evaluating
the effectiveness of automobile advertising. Silk and
Urban(1978) have purchaseintentionsas one inputfor
a new productmodel. Bass, Pessemier, and Lehmann
(1972) use purchaseintentionsin theirwell-known soft
drinkstudy. Sewall (1978) uses purchaseintentionsto
segment marketsfor proposednew (redesigned)prod-
ucts. In spite of this heavy use of purchaseintentions
data, there has been little effort to constructa formal
mathematicalmodel that establishes the intermediate
links between stated purchase intentions and actual
purchasebehavior.This paperis an attemptat building
such a model which will make purchase intentions a
better understoodand more useful concept.
First, a general frameworkfor analyzing purchase
intentionsdata is presented. This model focuses atten-
tion on three very different phenomenathat often are
implicitly and confusingly combined. Then within this
overall frameworka specific, mathematicallytractable
and statistically testable model is developed. Some
actualdatafrom the well-known studyby Juster(1966)
is analyzed.
The data used are purchase intentions for au-
tomobiles and household appliances. The data fit the
Purchase Intentionsand Purchase Behavior/ 65
model quite well, but this is not the importantpoint.
Rather the model highlights the differences between
intentions data for a majorpurchaseof an automobile
and a still substantial, but less major purchase of an
appliance.
In the automobile study, only 53% of those who
said that they were certain to buy a new automobile
actually did so. How can we assess this statistic? By
itself not muchcan be said;however, withinthe context
of our model the 53% statistic appears to be quite
"good." We will now proceed to develop the model
that allows us to make this evaluation.
The General Model
The overall model uses a three-steptransformationto
transformstatedintention,Is, into a purchaseprobabil-
ity estimate, p.
Step 1. Stated intention, Is, is transformedinto an
estimate of the true intention, It, by the TrueIntention
Model (TIM). This procedurehas a long history in the
psychometricliteraturewherean observedscore on, for
example, a multiplechoice test is used to estimate the
subject's true ability. The driving force for all of this
work is that the observed score may be written as
Observed Score = True Ability + Error (1)
The obvious analog for intentions is
Stated Intention = True Intention + Error
This literatureis well summarizedin the book by
Lord and Novick (1968) and articles by Lord (1959;
1965), Keats and Lord (1962), and Duncan (1974). A
less mathematicallydemandingarticlewhich illustrates
the basic notion can be found in Morrison (1973).
Briefly expressed, the TIM deflates the high stated
intentions and inflates the low stated intentions in an
effort to more accuratelyrepresenttrueintentionat the
time of the survey. This type of adjustment,known as
"regression to the mean" is requiredin all situations
where the observedphenomenacan be expressedqual-
itatively by (1). This regressionto the mean effect will
be quantitativelydeveloped in (7) within the context of
a specific TIM.
Step 2. The estimated true intention, It, is trans-
formed into an unadjusted purchase probability esti-
mate, •p, by the Exogenous Events Model (EEM). An
individual may not expect to buy a new appliance
duringthe next year andaccuratelystatethis fact on the
survey. However, six months later the dishwasher
motormay burnout leaving the individualwith a buy or
repairdecision. At this point in time he or she may elect
to buy. Similarly, an individualmay trulyplan to buy a
new car during the next year. However, a subsequent
66 / Journalof Marketing,Spring 1979
This content downloaded from 62.197.243.108 on Fri, 23 May 2014 10:35:51 AM
All use subject to JSTOR Terms and Conditions
sudden loss of income may make such a purchase
infeasible. The EEM allows for such changes in
circumstances-albeit in a very simple mannerin the
specific model.
Step 3. The estimatedunadjustedpurchaseprobabil-
ity,Iu, is transformedinto the estimated purchase
probability, , by the Probability AdjustmentModel
(PAM). There is no reason to believe that the average
statedpurchaseintentionwill equal the proportionthat
actually buy the productfor all productsand all time
frames. Hence, some transformationwill be needed to
adjustfor this bias. Thereare manypossible adjustment
formulas and, again in our specific model, we use an
extremely simple adjustment.
The general model can be summarizedby
Stated Intention
TIM
True Intention
EEM
UnadjustedPurchaseProbability
PAM
Purchase Probability
The readermay have noticed the first letter of each of
the two intentionsand two probabilities.However, we
will resist the temptation to call this the Three Step
STUP Model.
Each of the threetransformation models will now be
given a well-defined and quite simple mathematical
structure.The resulting aggregate model will then be
fitted to some of Juster's data. The research implica-
tions of this approach will be discussed in the final
section.
A Specific Model
A Beta Binomial True Intentions Model
Probablythe simplestmodel one could constructwould
be a linear relation between stated intention, Is, and
average true intention, It.
Namely,
It =a+k Is. (2)
The regressionto the meannotion impliesa > 0, k > 0,
anda + k <1. Graphically,this implies that(withIt on
the vertical axis) equation (2) is a straightline with a
positive interceptand a slope of less than 1, or equiva-
lently, the straightline makes an angle of less than 45
degrees with the horizontalaxis.
We could leave the model at this level of justifica-
tion (i.e., mere simplicity), but there is, in fact, good
reason for such a formulation.At the individuallevel,
the stated intentionis a randomvariable-recall equa-
tion (1)-with a meanat or nearthe trueintention.If the
individual is respondingon an n +1 point scale which
we can label 0,1,2, . . . ,n, then the most natural
individualmodel will be thatIs is distributedaccording
to the binomial distributionwith parametersn and It.
More formally,
P(Is
= r; n, It) = (n)
I(1-It) n-r,
r = 0,1 .... .n.
An underlyingand untestablemodel thatgenerates(3)
is thatan individualwith trueintentionItresponds1 or 0
in an independentfashion to each point on the scale
with probabilitiesIt and 1 - It, respectively. The stated
intentionthen is the sum of these 0,1 responses (i.e., a
binomial randomvariable). This process that yields a
binomially distributed stated intention may sound a
little strange, but it is in the same spiritof many of the
mathematicalpsychology models of attitudeformation
where subjectspick up cues in a probabilisticmanner.
In any event, we are not interestedin the detailedmicro
process of forming the stated intentions. Rather, we
only need (3) to be a reasonablyspecific representation
of the generalmodel: StatedIntention= TrueIntention
+ Error.
No direct evidence for or against (3) as a model of
stated purchaseintentionis in the literature,but Juster
(1966) does give some data that support the within
individual variabilitythat the binomial model yields.
He cross classifies individualsby their responses on a
10 point intention probability scale and a five point
"definitely will" to "definitely will not" scale (p.
677). The individualsshow sufficient inconsistencyto
make our binomialmodel at least a good startingpoint.
More will be said on this issue in the Discussion sec-
tion.
Whatever individual stated intentions model is
used, it is clear that true intention will vary across the
population of respondents. Letting g (10 represent the
probability density function of It, then the aggregate
distribution of is well-known mixed binomial
Is
1
= r; n, It)g(It)dl',
P(Is - r)=fP(Is
0
r= 0,1,. . . ,n.
There is a vast statistical and psychometric literature on
this general mixed binomial model (4). The first n
This content downloaded from 62.197.243.108 on Fri, 23 May 2014 10:35:51 AM
All use subject to JSTOR Terms and Conditions
moments of the unobservabledistributiong (It) can be
estimated from the first n moments of the observable
distributionon Is. (See Skellam 1948; Keats and Lord
1962.) An estimateof It fromIs can be obtainedfor this
general mixturemodel. (See Lordand Stocking 1976.)
However, for our purposes we will restrictg (It) to
being a beta distribution
=
g(It;a,/3) FF (a) F(8) (/3) It
(a+p) (5)
(1-/t)
0 < It < 1.
The beta is a flexible distributionthat takes on uni-
modal, U, J, andreverseJ shapes. Using (5), the mixed
binomial (4) becomes the well-known beta binomial
(3) model
P(Is = r; n,a,p) = (n) B(a + r,f3+n-r) (6)
r =
B(,/3)
0,1,... n.
where B
(a,p/) F (a) F ().
F (a+p)
This beta binomialmodel has been used extensively
in branchswitchingmodels(e.g., KalwaniandMorrison
1977) and in advertisingreach and frequencymodels
(e.g., Greene 1970). Furthermore,(6) is the unique
aggregatedistributionthat (assuminga binomiallydis-
tributedIs at the individual level) will yield a linear
relationshipbetween the expected value It andIs. This
is proved in Keats and Lord (1962) where they referto
the beta binomial as the negative hypergeometricdis-
tribution.
If we divide the statedintentionby n (so it now takes
on values 0, 1/n, 2/n, ... ,1), the linear relation shown
in (2) becomes1
= a +(7)
For the beta distribution(5) the parametersa, /3 >
0. Hence, when a+/3 approaches0, the slope of (7)
approaches1. That is, stated intentionis very close to
true intention. This is intuitively appealing because
when a+/3 -- 0 the beta distributionhas virtuallyall of
its mass near 0 or 1. For most people the binomial
variancein (3) will be near 0. Hence, the stated inten-
tion is very close to the true intention.
At the other extreme, when a +/3 becomes large the
slope in (7) approaches 0. This implies that stated
intention has very little relationship to true intention.
(4)
'Originallywe have the statedpurchaseintentiontake on the n + 1 integer
values between 0 andn. This allows us to make the binomialassumptionat
the individual level. However, when we compare stated intentions with
actualpurchaseprobabilitiesit is naturalto transformthese statedintentions
into the interval0 and 1. Therefore,the transformedstatedintentionswhich
takeon values 0, 1in, 2/n, ...,. 1 arenot strictlybinomiallydistributedat the
individuallevel since the range of this randomvariableis no longer 0, 1, 2,
Purchase Intentionsand Purchase Behavior/ 67
Again this is intuitivelyappealingsince the beta distri- no change occurs) and a/(a +/3), the expected value of
bution becomes a "spike" at its mean a/(a+p) when It, if a change does occur. This results in
a + /3-?o. Hence, all individualshavevirtuallythe same
true intention. The variationin stated intention is all . p
(-1-)c + (8)
"noise"-i.e., the "Error" component in (1)-and Pu - a + a+p+ a1 +,+1
accordinglythe true intentionestimateis not influenced = p E(It) + (1 - p)lt.
by the observed stated intention.
A detaileddiscussionof these two special cases and Note thatthe slope in (7) is reducedby a factorof I -p.
situations in between can be found in Sabavala and Whenp = 0, thereis no changeand(8) reducesto (7) as
Morrison(1977). In most purchaseintention settings, it should. Whenp = 1, everyone changes so that stated
the case of a small value of a +/3 will prevail. This will intentionat the time of the surveyhas no effect on actual
lead to a steep slope for a graph of true purchase purchaseprobability.
intentionversus statedpurchaseintention.Notice in (7) Since the probabilityof change p is independentof
when a+p3 = 1/2the slope of the graphis 1/ (/2 + 1) =
Is, thepopulationaverageforpawill equalthepopula-
2/3. Equivalently, this straightline makes a 34 degree tion averageoflt which is in turnthe populationaverage
angle with the horizontal axis. The Juster data yield of Is (rescaled to take on values 0,1/n ,2/n, . .. ,1).
values of a+P3 in the neighborhoodof ?2. This average is, of course, the average of the beta
However, when the graph uses actual purchase distribution(5)-namely, al/(a+p).
probabilities on the vertical axis instead of the esti- If there does exist some systematicbias in It (i.e..
mated It, a much flatter curve usually results. (See on average it over or understatesp, the average actual
Figures 1-4 which will be discussed when Juster'sdata purchaseprobability),then some form of adjustmentto
are analyzed.) If It is really a good estimate of the true (8) is required. This brings us to Step 3.
intention, why is there this apparentdiscrepancy be-
tween true intentionand actualbehavior?This leads us A Constant Bias Adjustment Model
to Step 2. The overall bias in (8) will be
A New Draw Exogenous Events Model b = Is - p (9)
As mentionedin the introduction,manythingscan hap- where Is is the average stated intention and p is the
pen to change an individual's purchaseintentionsdur- proportionthat actually purchase.
ing the time frameof interest. These exogenous events Again, there are numerous ways to adjust (8).
could be modeledin a varietyof ways. In particular,we However, with simplicity in mind, we merely subtract
could segment the individualsby theirstatedintentions b fromeach estimatePuyielding a purchaseprobability
and model each class separately. However, it seems estimate for each individual of
much more appropriateto begin with a simple model.
The spiritof this model will be to obtaina graphof
Pu, the unadjustedpurchaseprobability,versusIs where p =Pu - b,
this graphhas a smallerslope thanItversusIs. If we also
can retain a linear relation, so much the better. or
A model that accomplishes this reduced slope,
linear relationshipis as follows:
+ (I;
-
j
pc Is - b? (10)
(1
* With probability 1-p the individual has no
ii=pa
a+p p)Oa +(lP
a+P+1 , 1--p
a+)3+1 ),
change in true intentionand this true intentionis
in fact the unadjustedpurchaseprobability. While (8) producesestimatesthatare constrainedto lie
in the acceptable range of 0 to 1, (10) can yield values of
* With probability p the individual has a change in
a outside of this range. However, this will only happen
true intention. When this change occurs a new when ca, /3, and p are small (i.e., initial true intentions
value for It is drawn from the original beta distri- are concentrated near either 0 or 1 and few changes take
bution (5). This new value It then becomes the
place) and/or b, the bias, is very large. This disadvan-
unadjusted purchase probability. tage of (10) is balanced by the retained linearity be-
Basically, all this model is saying is that when a tween p and Is.
change occurs it is as if the individual is being replaced
by a randomly chosen, new individual from a popula-
tion withIt distributed according to (5). This model will Some Empirical Examples
yield apu which is a simple weighted average of (7) (It if The Juster (1966) paper is one of the few published

68 / Journalof Marketing,Spring 1979

This content downloaded from 62.197.243.108 on Fri, 23 May 2014 10:35:51 AM

All use subject to JSTOR Terms and Conditions
works where both purchase intention and actual pur- Probably (7 in 10) 7
chase behaviorhave been obtainedon the same sample Good possibility (6 in 10) 6
of individuals. For both automobiles and household Fairly good possibility (5 in 10) 5
appliances, individuals responded to the following Fair possibility (4 in 10) 4
question: Some possibility (3 in 10) 3
Takingeverythinginto account,whataretheprospects Slight possibility (2 in 10) 2
thatsomememberof yourfamilywill buya- sometime Very slight possibility (1 in 10) 1
during the next - months; between now and --? No chance, almost no chance (1 in 100) 0
The time frames were six, 12, and 24 months for both
automobiles and appliances. The respondentchose a Juster then calculates the actual purchaseprobability
numberon a flash card which containedthe following for each intentions class. However, because of rela-
statements. tively small samplesizes he only presentsthe resultsfor
the following five aggregateintention groups:
Certain, practicallycertain (99 in 100) 10
Almost sure (9 in 10) 10, {9, 8, 7}, {6, 5, 4}, {3, 2, 1}, 0.
9
Very probably(8 in 10) 8 Juster's actual purchases were in a six month

TABLE 1
Automobiles: 6 and 12 Month Intentions vs. 6 Month Purchases

IntentionScale No. with 6 Month IntentionGroup Purchase

Intention PurchaseProbability Probability
Sample Size
0.0 345 (347) .10 300
0.1 29 (27)-
0.2 16 (15) .30 53
0.3 14 (11)J
0.4 3
0.5 (9)]
10 (7) .65 17
0.6 6 (7)J
0.7 6 (6)
0.8 6 (6) .36 14
0.9 5 (6)J
1.0 11 (9) .55 11
Total:451 Mean:.17 395

a = .074, = .67, / = .099, p = .37, b = -.071

IntentionScale No. with 12 Month IntentionGroup Purchase

Intention PurchaseProbability Probability
Sample Size
0.0 293 (288) .07 256
0.1 26 (35)
0.2 21 (20) .19 57
0.3 21 (15)J
0.4 10 (13)
0.5 9 (11) .41 32
0.6 12 (11)J
0.7 13 (10)
0.8 11 (11) .48 31
0.9 10 (12)J
1.0 21 (20) .53 19
Total:447 Mean: .17 395

a=.12, 1,
= .56 /,=.17 p =.22, b= 0

Purchase Intentionsand Purchase Behavior/ 69

This content downloaded from 62.197.243.108 on Fri, 23 May 2014 10:35:51 AM

All use subject to JSTOR Terms and Conditions
 TABLE2
Appliances: 12 and 24 Month Intentions vs. 6 Month Purchases
IntentionScale No. with 12 Month IntentionGroup Purchase
Intention PurchaseProbability Probability
Sample Size
0.0 2377 (2386) .017 2111
0.1 87 (84)1
0.2 57 (45) .053 150
0.3 29 (32)J
0.4 23 (25)
0.5 22 (21) .111 63
0.6 21 (19)
0.7 14 (18)]
0.8 11 (17) .184 38
0.9 17 (18)J
1.0 30 (24) .105 19
Total:2688 Mean:.026 2381
e= .034, = .69,
Is =
.047, p = .79, b = .021

IntentionScale No. with 24 Month IntentionGroup Purchase

Intention PurchaseProbability Probability
Sample Size
0.0 2174 (2177) .015 1429
0.1 95 (115)]
0.2 99 (64) .034 206
0.3 41 (46)J
0.4 38 (38)
0.5 28 (34) .045 206
0.6 30 (31)J
0.7 29 (31)]
0.8 36 (32) .084 83
0.9 24 (39)J
1.0 80 (69) .215 65
Total: 2674 Mean:.026 1989
a=0.05, /3= .51, IS= 0.90, p = .74, b = .064

period-December 1964 to June 1965. Yet, he used
these same purchasedata for assessing six, 12, and 24
monthintentions.Whenthe 12 and24 monthintentions
are compared to six month purchases we seem to be
comparing apples to oranges. However, as the au-
tomobile data will show there is some advantage to
doing these differing time frame comparisons.
Table 1 gives the results for automobiles. The
sample histogramfor both six and 12 month intentions
arepresented.The numbersin parenthesesarethe fitted
beta binomial expected values (roundedto the nearest
integer). The beta binomial parametersa and /3 are
estimated by the method of moments while p is esti-
2Scalingthe statedintentions0, 1 10 the samplemean,X, andsample
variance.S2, of these intentionsare....
calculated. We then solve for the values
a and /3 that make the theoreticalmean 10l/(a+/P)=X and the theoretical
variance lOaJp(a+p+ 1O)/(a+)2(a+p+ 1)=S2. The theoreticalslope of
actual purchase probability versus stated intentions is (l-p)/(a+P3+1).
Since we previously calculated a and 8 we now find p which makes the
observedslope equal (l-p)/(a++3 1).
matedby a simple regressionof intentionversus actual
purchaseprobabilityfor the intentiongroup.2When the
S1, 2, and 3} groupwas used, the middlevalueof 2 was
the value used in the regression,and similarlyfor {4, 5,
and 6} and {7, 8, and 9}. The bias is, of course, the
difference between the mean intention and the total
sample purchaseproportion.Finally, some of the pur-
chase data were not collected so that the two sample
sizes differ somewhat. Justerdoesn't explain why he
kept some of the intentiondatathathad no correspond-
ing purchase data. In any event we had no way of
correcting this difference.
Table 2 contains analogous results for six house-
hold appliances (air conditioners, refrigerators,wash-
ing machines, clothes dryers, television sets, and dish-
washers). Juster does not present data for the indi-
vidual appliances. That is, each individualrepresents
six datapoints-an intention, purchasepair for each of
the six appliances. In addition, only 12 and 24 month

70 / Journal of Marketing, Spring 1979

This content downloaded from 62.197.243.108 on Fri, 23 May 2014 10:35:51 AM

All use subject to JSTOR Terms and Conditions
intentions were collected. Of course, they are both
being comparedto the six month purchasedata.
Since we are grouping all six appliancestogether,
the resultingparametervalues arean approximateaver-
age across all six appliances. We could makethe rather
heroic assumption of common parameters for each
appliance, but that would be stretchinga little too far.
Therefore,we will use the averageparameterinterpre-
tation. The next section will show that the behaviorof
the purchase intentions for this "average appliance"
differs considerably from the behavior of the au-
tomobile purchase intentions.
Figures 1-4 plot the stated purchase intentions
versus the intentiongroup purchaseprobabilities.The
straightline in each graph representsequation (10).
Summary of Empirical Findings
A number of points emerge from Tables 1 and 2,
Figures 1-4, and the summaryresults in Table 3.
* The beta binomial fits the stated intentions his-
togram quite well in all four cases. This only
implies that the composite hypothesis of indi-
vidual binomial responses and beta distributed
true intentions is a good approximation.
Nevertheless, these results are encouraging for
the beta binomial true intentionsmodel.
* The six monthautomobileintentionsdataare not
very good. They exhibit a large downwardbias
and they are not at all monotonicallyrelated to
actual purchaseprobabilities.
* The 12 month automobile intentions fit the six
monthpurchasedata very well. There is no bias
and the intentionsare linearlyrelatedto purchase
probabilities.
* Both 12 and 24 month applianceintentionshave
the positive bias that is expected. The purchases
only cover a six monthperiod-leaving six or 18
months in which to fulfill the intentionby actu-
ally buying.
TABLE3
SummaryStatistics
Category Time a 1
/
Frame
Automobile 6 months .074 .67
Automobile 12 months .120 .56
Appliances 12 months .034
.691
Appliances 24 months .050 .51
This content downloaded from 62.197.243.108 on Fri, 23 May 2014 10:35:51 AM
All use subject to JSTOR Terms and Conditions
* Both appliancefigures show a much flatterrela-
tionship between actualpurchasesand stated in-
tentions. This implies a higherp value (see Table
3) which in turn implies a greaterinfluence of
unexpected exogenous effects.
* The quantity 1/(1+ca+) is an index that varies
between 1 (when ac + 83-40 and all of the true
intentionsare concentratedat the extreme values
of 0 and 1) and 0 (when a~+ -/34 and all of the
true intentions are the same). Hence,
1/(1+a+ f) can be called a polarizationindex of
true intentions.In all four cases, the polarization
is about the same. See Sabavala and Morrison
(1977) for a more detailed description of this
index.
* The parameterestimates for a and f are quite
good since large sample sizes are involved. The
estimate of p is somewhat more suspect since
only five points(albeiteach pointis an averageof
a large numberof other datapoints) are used. In
addition, both variables are subject to error so
that the usual least squareassumptionsdon't ap-
ply. Nevertheless, the slopes for the automobile
curves are clearly much greaterthanthose of the
appliancecurves.
Discussion
The three-stepmodel presentedhere focuses attention
on a numberof importantissues that tend to be over-
looked when statedpurchaseintentionsarecomparedto
actual purchase behavior. Our model forces the re-
searcherto explicitly consider:
* The numberof scale points
* The numericalvalue of each point on the scale
* The adjectives associated with each point on the
scale
* Whether a direct probability response scale
should be used
* The reliability of stated intentions
p b
1+a+P
.57-.071
.60 0
.58 .021
.64 .79 1 .064
.74
Purchase Intentions and Purchase Behavior / 71
 FIGURE1
Table 1 shows that 12 month intentions are better
Automobiles:6 Month Intentions (both in terms of bias and monotonicity)in predicting
1.0 six month automobile purchases than are the corre-
spondingsix monthintentions.This may be an isolated
case, but it is a concept worthstudyingfurther.Perhaps
betterestimates of actualpurchasesover x monthscan
be obtainedby having statedintentionsover y months.
Clearly, automobileintentionsaremore stableover
time than applianceintentions. This would lead one to
0
conjecture that purchase intentions for low cost, low
L-
involvement items are even less stable-yielding an
even smaller slope in the purchaseprobabilityversus
purchase intentions graphs. However, since the time
0r1 framefor these low cost productintentionsdatais often
two weeks, the slopes are not comparableunless some
adjustmentis made for the differing time scales.
The model also directs attention to the "yea
sayer"-"nea sayer" issue. Clearly, the absolutepur-
chase probabilityfor the "certain to purchase" group
0 tells us nothing. Figure3 shows thatthis groupactually
0 1.0
Is purchases less than the .8 intention group. This non-
monotonicity implies a "yea sayer" effect, but the
FIGURE2
otherthreegraphsdo not indicatesuch an effect. Recall
Automobiles:12 MonthIntentions
that in the introduction we posed the question of
1.0
whether a 53% purchaseprobabilityfor a "definitely
will buy" group was a "good" value. We really can't
say for sure, but by looking at Figure 2 we do see that
this 53% value is "consistent" with the model. Hence,
it is a "good" value.
Ourmodel suggests some naturalways of segment-
ing intentions data into meaningful categories. When
sufficient data have been gathered, we could separate
o productsalong the following dimensions.
* Reliability of stated intentions
* Slope of ~vs. Is (i.e.., the stabilityof true inten-
tions)
* Linearityof p vs. I
* Differencebetweenp and Is (i.e., the bias in the
_
stated intentions)
* The most appropriatetime frames
0 rs 1.0
Otherdimensionscome to mind, but again this list will
suffice.
* The transformationfrom statedintentionsto true
We need to make some comments on empirically
intentions
* The distribution of true intentions across the assessing the model. Most data sets will contain only
stated intentionsand actual purchasebehavior.3There-
populationof consumers
* The appropriatetime frame for the intentions
* The effect of "unexpected" events over the time 3At the individual level, we could test the model by correlatingthe esti-
mated purchaseprobabilityp with the actual binary purchasevariable(the
frame individual did or did not purchase) across all individuals in the sample.
* Systematicdiscrepenciesbetween averageinten- However, the resulting empiricalcorrelationwould need to be interpreted
tions and average purchaseprobabilities carefully.A "low" value of r (or r2)may not be quite so low. (see Morrison
1972.) Alternatively,we might wish to classify the individualsas buyersor
This list could go on, but the 10 items mentioned will nonbuyersbased on stated intentions. Again the resulting2 x 2 classifica-
tion table would need to be interpretedwith care. (See Beckwith and
suffice. Morrison 1977.)

72 / Journalot Marketing,Spring 1979

This content downloaded from 62.197.243.108 on Fri, 23 May 2014 10:35:51 AM

All use subject to JSTOR Terms and Conditions
fore, the three steps of the model are not tested individ- FIGURE3
ually. (Studies aimed at determiningthe reliabilityand Appliances: 12 MonthIntentions
stability of purchaseintentionswill help determinethe 1.0
validity of the first two steps in the model.) Any Juster
type data where purchaseprobabilitiesare linearly re-
lated to stated purchaseintentions are consistent with
our model if the statedpurchaseintentionshave a beta
binomial distribution. The advantage of the formal
three-step model is that the resulting parametersde-
compose the empirical slope of this linear relationship
into the componentdue to the heterogeneityof the true o

purchase intentions (determinedby a and 3) and that ,,,

due to the propensityof individualpurchaseintentions -r-
m0
to change (determinedby p).
m0

Finally, we must discuss the key issue of parameter

stationarityover time. The beta distributionparameters
a and / are estimated from the histogram of stated
intentionsso they pose no stationarityproblem.Thatis, a
0 Is
they are estimated at the currentpoint in time. How- 1.0

ever, in projectingthese intentionsinto actualpurchas-

ing behavior, p, the stability of the individual true FIGURE4

intentions,and b, the bias, are obviously crucial. For a Appliances:24 MonthIntentions

new productthe problemis even more severe. Namely, 1.0
we have to first ascertainthe appropriateproductcate-
gory and/oridentifypast new productsthatwere similar
to the proposed new product. Then, the parameters
estimated for these appropriatepast productsmust be
assumedvalid for the currentnew productso thatequa-
tion (10) type forecasts can be made. >r-

At the moment there is not an adequateset of pur-

chase intentions with follow up purchase behavior
studies in the literatureto make the above type of
forecast. Hopefully, this paper will supply a unifying
frameworkfor the design of such studies and that the
researchersinvolved will be motivatedto publish their
findings. If this is done the probabilityis very high that
empirical regularitieswill emerge that will make pur-
chase intentionsdata much more useful to the market- a0
0 Is 1.0
ing community.

REFERENCES Multiple Choice Tests in the MisinformationModel," Journal

Axelrod, J.N. (1968), "Attitude MeasuresthatPredictPurchase,"9
Journal of AdvertisingResearch, 8 (March), 3-18.
Bass, F.M., E.A. Pessemier, and D.R. Lehmann (1972), "An
Experimental Study of the Relationship Between Attitudes,
Brand Preferences and Choice," Behavioral Science, 17
(November), 532-541.
Beckwith, N.E. and D.G. Morrison(1977), "StochasticInterpreta-
tion of 2x2 Classification Tables," Journal of the American
Statistical Association, 72 (June), 303-308.
Duncan, G.T. (1974), "An EmpiricalBayes Approachto Scoring
of the American Statistical Association, 69 (March), 50-57.
Greene, J.D. (1970), "Personal Media Probabilities,"Journal of
AdvertisingResearch, 5 (October), 12-18.
Juster, F.T. (1966), "Consumer Buying Intentionsand Purchase
Probability:An Experimentin Survey Design," Journal of the
American Statistical Association, 61 (September), 658-696.
Kalwani, M.U. and D.G. Morrison(1977), "A ParsimoniousDe-
scription of the Hendry System," Management Science, 23
(January),467-477.
Keats, J.A. andF.M. Lord(1962), "A TheoreticalDistributionfor
Mental Test Scores," Psychometrika, 27 (March), 59-72.

Purchase Intentionsand Purchase Behavior/ 73

This content downloaded from 62.197.243.108 on Fri, 23 May 2014 10:35:51 AM

All use subject to JSTOR Terms and Conditions
Lord, F.M. (1959), "An Approach to Mental Test Theory,"
Psychometrika, 24 (December), 283-302.
(1965), "A Strong True-Score Theory with Applica-
tions," Psychometrika, 30 (September), 239-270.
and M.R. Novick (1968), Statistical Theoriesof Mental
Test Scores, with contributionsby Allen Birnbaum. Reading,
MA: Addison-Wesley.
and M.L. Stocking (1976), "An IntervalEstimate for
Making StatisticalInferencesabout True Scores," Psychomet-
rika, 41 (March), 79-87.
Morrison,D.G. (1972), "Upper Bounds for CorrelationsBetween
BinaryOutcomesand ProbabilisticPredictions,"Journal of the
American Statistical Association, 67 (March), 68-70.
(1973), "Reliability of Tests: A Technique Using the
'Regression to the Mean' Fallacy," Journal of MarketingRe-
search, 10 (February),91-93.
Sabavala, D.J. and D.G. Morrison(1977), "A Model of TV Show
Loyalty," Journal of Advertising Research, 17 (December),
35-43.
Sewall, A. (1978), "'MarketSegmentation Based on Ratings of
ProposedProductDesigns," Journalof MarketingResearch, 15
(November), 557-564.
Silk, A.J. andG.L. Urban(1978), "Pre-Test-MarketEvaluationof
New Packaged Goods: A Model and MeasurementMethodol-
ogy," Journal of MarketingResearch, 15 (May), 171-191.
Skellam, J.G. (1948), "A ProbabilityDistributionDerivedfrom the
Binomial Distributionby Regardingthe Probabilityof Success
as a Variable Between Sets of Trials," Journal of the Royal
Statistical Association, B 10, 257-265.
Smith, G. (1965), "How GM MeasuresAd Effectiveness," Print-
er's Ink, (May 14), 19-29.

74 / Journalof Marketing,Spring 1979

This content downloaded from 62.197.243.108 on Fri, 23 May 2014 10:35:51 AM

All use subject to JSTOR Terms and Conditions